3D printed designer color routers with low refractive index for low-light imaging

Abstract

Color imaging systems that underpin modern technologies still suffer in low-light conditions because of the inherently lossy filtering mechanism of conventional Bayer color filters. Although Bayer-type filters integrated with complementary metal–oxide–semiconductor sensors remain the dominant architecture, they typically transmit only about 30% of incident white light, leading to noisy images under weak illumination. Color routers offer an alternative by redirecting different spectral components of light to corresponding photodiodes in the near field. While recent color routers based on high-refractive-index metasurfaces have shown promise, low-refractive-index materials remain largely unexplored despite their potential for higher transmittance and cost-effective mass production. Here, we demonstrate inverse-designed, two-photon-polymerization 3D-printed pixel-level color routers using low-index IP-L resin (n ≈ 1.5). The color routers achieve transmittance of ~87% in simulation and ~70% experimentally across the visible spectrum. Monochrome-sensor imaging further confirms improved color fidelity and shorter exposure times, highlighting their potential for next-generation high-efficiency imaging platforms.

The authors demonstrate a 3D-printed low-refractive-index color router that replaces absorptive color filters, improving light collection and enabling brighter, more efficient color imaging under both normal and low-light conditions.

Introduction

Optical sensors that integrate various wavelength-selective filters with broadband photodiodes capture multi-spectral information, which enabling devices such as spectrometers and imaging systems operating across diverse spectral ranges operating across diverse spectral ranges^1–7. Among these, color imaging sensors are the most widely employed, supporting applications from digital photography and medical diagnostics to autonomous navigation and AR/VR, where the accurate capture of spatial and spectral information is fundamental to visual data processing. The performance of a color imaging system is governed by its spatial resolution, spectral fidelity, and light-utilization efficiency, all of which become increasingly critical as device footprints shrink and application demands continue to expand.

In modern imaging systems, light forming an image through the optical setup is captured by an image sensor consisting of an array of discrete pixels, where photodetectors-most commonly complementary metal–oxide–semiconductor (CMOS) sensors-convert incident photons into electrical signals. For color imaging, conventional CMOS sensors integrate color filters (CFs) with photodetectors to capture spectral information. Motivated by the trichromatic nature of human vision, these systems generally sample light in three spectral passbands: red (R), green (G), and blue (B). The most widely adopted configuration is the Bayer filter (BF) array, in which a 2 × 2-unit cell comprises two green, one red, and one blue subpixels (RGGB) to balance spatial resolution and sensor fill factor.

Recent developments in high-transmission CFs have delivered peak efficiencies within the passband nearing 90%^4,8–10. However, despite this impressive transmittance at the subpixel level, substantial optical energy is lost across the pixel unit. For instance, under monochromatic red illumination, the light incident on the green and blue subpixels within a Bayer unit is entirely absorbed by the filter (Supplementary Note 1, Fig. S1). As the demand for compact imaging systems continues to grow, pixel sizes are being reduced to increase spatial resolution. Yet, this miniaturization exacerbates optical crosstalk between adjacent subpixels, degrading both color fidelity and image quality. Moreover, the intrinsic filtering mechanism fundamentally limits performance in low-light conditions, which is conventionally compensated by extending exposure times. Prolonged exposure, in turn, hampers image acquisition time and reduces image accuracy in some scenarios¹¹.

The concept of CR^12,13 is inspired by the spectral sorting functions afforded by multiwavelength diffractive optics^14,15 and subwavelength nanophotonic devices^6,16. In contrast to the subtractive nature of conventional CFs, which absorb light outside the passband, CRs selectively direct the respective desired wavelength bands to predefined spatial locations at the unit-cell level, thereby potentially improving light-utilization and transmittance efficiency (Fig. 1a, b). To overcome the inherent limitations of conventional CF architectures, several studies have also explored integrating nanoantenna resonators directly with photodiodes¹⁷. Despite the modest responsivity of the photodiode architecture, doing so mitigates optical crosstalk induced by oblique illumination in reduced photodetector footprints.

Fig. 1. Comparison of traditional Bayer color filter and designer color router.

a Schematic of the traditional microlens, Bayer filter, and photodetector arrays where different channel of light is filtered into subpixels within a single unit-pixel, and (b) Schematic of designer 3D-printed CR consisting of nanopillars of varying sizes in transparent low-refractive-index IP-L resin working collectively to channel light to their respective subpixels. c Plot of refractive index and chromatic dispersion for a selection of materials commonly used in optical and photonic applications.

Considerable progress has been made in realizing high-efficiency CRs using HRI metasurfaces^18–28. Nano-antennas fabricated from HRI materials such as TiO₂, GaN, or SiNₓ provide strong optical confinement and powerful light-manipulation capability, enabling efficient Bayer-type spectral routing in compact pixel architectures. However, the polarization sensitivity and limited transmission efficiency are longstanding challenges^{12,13,29–31}. Although several recent CR studies have demonstrated good performance in four-pillar macro-pixel configurations^32,33, these devices generally rely on tightly controlled geometries and dimensions. As a result, their fabrication often requires substantial fine-tuning, including high-quality thin-film deposition, high-resolution lithography, and precise etching³⁴. Against this backdrop, the key advance of the present work is the realization of efficient color routing in a low-refractive-index (LRI) material platform without a complicated multi-step fabrication process. Such platforms have remained largely unexplored for CRs because their weaker optical confinement in the visible range is generally considered unfavorable for compact spectral routing (Fig. 1c). Here, the use of 2.5D architecture is also significant: rather than simply increasing the geometrical degree of freedom, it provides a more practical and repeatable fabrication route for LRI photonic structures, while preserving the essential vertical structural variation required for routing. This makes the proposed approach more compatible with scalable low-index nanophotonic fabrication and demonstrates that efficient Bayer-type CR operation is not limited to HRI materials.

LRI materials—such as optical glasses, fluorides, and polymers—offer higher transparency across the visible spectrum, extremely low optical losses, and notable advantages in cost-effectiveness and manufacturability over their HRI counterparts. Incidentally, many LRI materials possess low stiffness, enabling scalable mass-production techniques such as roll-to-plate³⁵ or nanoimprint lithography³⁶. Recent advances in high-resolution fabrication methods for LRI materials-including direct laser writing^37,38, nanoimprint, grayscale electron-beam lithography³⁹ and two-photon polymerization lithography (TPL)^40–42-have opened new opportunities for realizing complex nanostructures with subwavelength resolution. Importantly, fabricated LRI nanostructures can also serve as templates for casting into other materials, facilitating cost-effective and scalable manufacturing. Unlike HRI metasurfaces, which often require heterogeneous material systems and multi-step processing, LRI CRs can potentially be fabricated in a single step using a single material, crucially with greater tolerance to fabrication imperfections. By comparison, conventional dye-based CF fabrication typically involves multiple sequential steps to integrate the three RGB filters, combined with precise alignment procedures (Supplementary Note 2, Fig. S2). These considerations make LRI-based CRs a compelling direction for next-generation imaging systems-provided their intrinsic limitations in dispersion and light manipulation can be mitigated.

In this work, we employ TPL to directly prototype three-dimensional LRI CRs, leveraging its capability to produce high-aspect-ratio nanostructures with feature sizes down to ~150 nm. TPL has been successfully applied to a wide range of nanophotonic devices, including metalenses, diffractive optical elements (DOEs), and photonic crystals^40,43–48. This strategy enhances light control through high-resolution nano-structuring and 2.5D geometries, which introduce additional design degrees of freedom and are compatible with scalable casting-based manufacturing^49–51.

We design a 2.5D CR based on LRI materials. Using inverse design⁵², the CR geometry was optimized to achieve high-efficiency routing of light, yielding simulated peak routing efficiencies of 60%, 61%, and 62% for the blue, green, and red channels, respectively, with an overall average efficiency of 87%. The devices were fabricated via two-photon polymerization lithography using IP-L photoresist. Experimentally, the peak routing efficiencies were 40%, 51%, and 48% for the blue, green, and red channels, respectively, with an overall average efficiency of 70%. To test the color reconstruction performance, we compared images captured by a monochrome CMOS sensor equipped with our CR to those obtained from an identical commercial CMOS sensor equipped with standard RGB CFs. In contrast to CFs, our CRs physically redirect incident light into designated regions, retaining more optical energy at the unit-cell level. This advantage is particularly significant in low-light scenarios, where the CR experimentally achieves higher light-collection efficiency than conventional CF-based systems. We further quantified its performance under reduced illumination by reconstructing color at progressively lower incident illuminance, revealing promising performance under relatively low-light conditions, enabling improved performance in applications such as astrophotography, night photography, and other low-light environments.

Results

Color-router design and optimization

We designed a metasurface CR in 2.5D structures (which makes it replicable with casting methods, unlike 3D structures) on a square grid of RGGB unit cells. Each unit cell comprises four cylindrical nanopillars (Fig. 2a), with nanopillars working collectively to scatter the three colors to their respective directions. As shown in Fig. 2a, the pillars located beneath the red, green, and blue focal spots are hereafter referred to as red, green, and blue pillars, respectively. Here, we assign six geometric parameters, i.e., heights and radii for 3 nanopillars, with the fourth green nanopillar repeated due to the symmetry in the collection domain to simulate the RGGB behavior. Thus, with six geometric parameters and well-defined objective functions, we employed an inverse-design approach based on particle swarm optimization (PSO) to determine the optimal pillar geometries, as detailed in Supplementary Note 6. Our goal is to achieve efficient RGB color separation by routing incident red (600–700 nm), green (500–600 nm), and blue (400–500 nm) light to their respective photodetector regions, each measuring 1 μm square. To quantify performance, we define the objective function as the smallest of $T_{\lambda }$, i.e., $\min \left(T_{\lambda }\right)$, where $T_{\lambda }$ refers to peak efficiency in different channels of $T_R$, $T_G$ and $T_B$. The fitness function was not defined to minimize the spectral peak intensity itself. Rather, the objective was to maximize the minimum routing efficiency among the selected target channels/wavelengths. In our optimization, the efficiencies at the target RGB wavelengths were first extracted from the simulated structure, and the smallest of these values was used as the objective metric. This strategy was chosen to avoid solutions in which only one or two channels perform well while the weakest channel remains poor. In other words, the optimization was designed to improve the worst-performing target response and thereby obtain a more balanced overall color-routing performance across the three bands, rather than simply maximizing an individual peak. $T_{\lambda }$ is calculated by following the function:

$$\begin{array}{cc}\hfill T_{\lambda }=\frac{{\int }_{{\mathrm{A}}_{\lambda }}{\mathrm{I}}_{\lambda }\mathrm{d}x\mathrm{d}y}{{\int }_{{\mathrm{A}}_0}{\mathrm{I}}_{\lambda }\mathrm{d}x\mathrm{d}y}\hfill & \hfill \lambda =R,G\left(G1+G2\right),B\hfill \end{array}$$ 1

Fig. 2. Schematic of 2.5D color router and its performance.

a Schematic of the CR unit cell consisting of 4 cylindrical posts of varying heights $H_{1-3}$ and diameters, $D_{1-3}$. Pillars corresponding to the two routed green regions share the same dimensions from symmetry. Light traveling upward through the pillars gets routed to the R, G₁, G₂, and B regions at the designed focal plane. b Plot of simulated transmittance as a function of wavelength for the inverse-designed CR monitored at the focal plane. c Plot of experimentally measured transmittance as a function of wavelength for the inverse-designed CR monitored at the focal plane. d Top-view and tilted-view (inset) SEM images of the CR patterned using TPL in Nanoscribe IP-L resin. e Optical microscope image of the CR at the focal plane with white-light illumination. f Optical micrographs at the focal plane with red, green, and blue light illumination, respectively (i)–(iii), with good agreement with simulation results despite the visible mismatch between fabricated structures from intended design (iv)–(vi).

Here $T$ represents the transmission efficiency at the wavelength corresponding to red (635 nm), green (545 nm), and blue (455 nm) light. ${\mathrm{A}}_0$ denotes the entire unit area at the focal plane (2 × 2 ${\mathrm{\mu }\mathrm{m}}^2$ region, 5 μm above the top of the CR), while ${\mathrm{A}}_{\lambda }$ indicates the 1/4 region of the focal plane which represents the 1 × 1 ${\mathrm{\mu }\mathrm{m}}^2$ target light-collection region for each RGB channel. Accordingly, the nanopillars were optimized with parameters of height H_B = 1.2 μm and diameter D_B = 0.56 μm for the blue pillar, red nanopillar with H_R = 2.0 μm, D_R = 0.68 μm, and green nanopillars with H_G = 1.9 μm, D_G = 0.48 μm. To verify the efficacy of our CRs, we first evaluated the transmittance of the metasurface under white-light illumination (Fig. 2b) in silico. Simulated results show a peak transmittance of $T_{\lambda }$ ~ 60%, 61%, and 62% across the blue, green, and red channels. As mentioned previously, when the entire unit is considered as the input light, the results of the blue, green, and red regions demonstrate a high transmissivity, in contrast to the ~20% transmissivity $T_{\lambda }$ of commercially available Bayer filters (see Supplementary Note 1). For a single period, the CR achieves a transmissivity of ~87%, whereas the filter-based system drops to ~35%, highlighting the feasibility of CRs for higher light-utilization efficiency.

Using the optimized design parameters, we fabricated 1 × 1 ${\mathrm{mm}}^2$-sized metasurfaces consisting of the designed nanopillar arrays via TPL using a Nanoscribe Photonic GT system and IP-L resin (refractive index n = 1.52). We 3D-printed nanopillar structures with varying heights in a single lithographic step (details in the “Methods” section). A CR sample comprising 500 × 500-unit cells was successfully fabricated. The scanning electron microscopy (SEM) micrographs shown in Fig. 2d exhibit an agreement with the intended design, with deviations attributed to fabrication imperfections. The rings along the length of the cylinders are caused by the layer lines of TPL. Remarkably, the imperfections of the structure do not appear to affect the color-routing performance severely, as shown in Fig. 2d. The fabricated sample was subsequently used for optical characterization and performance evaluation of the CR.

To experimentally verify the transmission and routing performance, we performed an additional dedicated optical characterization using a supercontinuum light source (SuperK SELECT UV/nIR1, NKT Photonics). In the measurement setup, the broadband illumination was first attenuated by an attenuator and then directed onto the fabricated CR device. The transmitted optical signal was collected through a 100× objective lens with NA = 0.9 and recorded using a monochrome CMOS camera. For transmission calibration, a reference signal was collected under the same optical path but without the CR, and the transmittance was then calculated by integrating the intensity within the detection region of the CR and normalizing it to the corresponding reference intensity. In addition, the overall transmission spectrum was independently collected using a spectrometer by averaging over a 20 × 20-unit area. As shown in Fig. 2c, the experimentally measured peak routing efficiencies are approximately 40%, 51%, and 48% at 450 nm, 550 nm, and 675 nm, respectively. These values are lower than the simulated values, corresponding to an overall reduction, which can be reasonably attributed to fabrication-induced geometric deviations from the optimized design, material-property mismatch between the simulated and fabricated structures, surface roughness, as well as minor alignment and collection uncertainties in the optical measurement setup.

To assess the color-separation capabilities of the CR, we illuminated the sample with white light and imaged the focal plane under an optical microscope (Nikon Eclipse LV100ND, the 100× Nikon objective lens with NA of 0.9). As shown in Fig. 2e, the transmitted light forms a well-defined square array corresponding to the RGGB configuration, with distinct red, green, and blue spots. The red and blue channels exhibit circular beam profiles with diameters of approximately 0.65 μm, while the green channels display slightly elongated elliptical profiles with a length of about 0.85 μm. Despite these differences, all beam profiles are smaller than the typical pixel size (~1 μm) of commercial CMOS image sensors, indicating low spatial crosstalk in the CR design. The detailed light propagation video is included in Supplementary Video 1 and Supplementary Note 10. It clearly demonstrates the optical response of the 3D-printed low-index color router recorded by optical microscopy while the focal plane is scanned along the z-axis. As the imaging plane moves away from the structure, the initially unseparated light pattern gradually evolves into a clearly resolved RGGB color-routing pattern, demonstrating successful spatial separation of red, green, and blue light.

To further verify the routing accuracy of each subpixel type, we illuminated the CR with spectrally filtered light using bandpass filters at 458 nm, 532 nm, and 633 nm (FL457.9-10, FL532-10, and FL632.8-10 THORLABS). As shown in Fig. 2f(i–iii), blue and red light were routed to the center of the subpixel in each unit cell, respectively, while green light was directed along the sides of the square array. These experimental results closely match numerical simulations. Approximately 25% of the blue light was observed to leak into the red channel, likely due to fabrication imperfections and slight angular deviations in the incident beam (in our setup, the light source was a bulb with an incident angle range of about 10°). This issue could be mitigated by integrating the CR with a microlens array.

Design principle of CR

Having established the viability of LRI CRs, we next examined the physical principle by which the structure redistributes incident light within a unit cell into predefined subpixel positions. Because the routing function is defined at a focal plane 5 μm above the structure, the key requirement is controlled near-field redistribution of optical energy rather than spectral filtering alone. To realize this within a compact pixel footprint, subwavelength 2.5D nanopillars are employed as wavelength-dependent scattering elements whose geometry and arrangement together determine the field distribution at the detection plane.

In our previous work⁵³, we demonstrated that periodic 2.5D nanopillars are effective for manipulating broadband light distributions and generating structural color. As shown in Supplementary Note 4, we fabricated periodic nanopillar arrays with varying heights, which produced distinct color variations, thereby confirming their ability to modulate far-field spectral responses. Here, we extend this concept to the near-field regime and find that an individual nanopillar already exhibits position-dependent routing behavior at the focal plane. To investigate the near-field distribution, we analyzed the light at the focal plane positioned 5 μm above the CR plane (Fig. 3a). When a nanopillar is placed at the center of the unit, the resulting spectrum remains flat, consistent with the incident light, and the energy is uniformly distributed across all quadrants as well as the overall detection region. In contrast, when the nanopillar is displaced toward one quadrant (orange pillar 2 in Fig. 3a), the spectrum measured in that quarter region becomes position dependent, indicating that even a single pillar can perturb the local energy distribution at the focal plane. In this work, our objective is to utilize the light incident on the entire unit cell and redirect it to the desired subpixel positions. Although changing the position of a single nanopillar already demonstrates the ability to manipulate the local spectrum within a given quadrant in the near field, this is insufficient for independently optimizing three different color channels. Thus, four different nanopillars are strategically positioned beneath each focal spot, with the expectation that their mutual interactions will collectively redirect the three-color channels.

Fig. 3. Near-field and far-field performance of single, periodic single-type and color router nanopillars.

a The transmittance spectra of the full focal-plane region (2 μm × 2 μm) and the subpixel region (1 μm × 1 μm, corresponding to one-quarter of the focal plane) were collected at a plane located 5 μm above the substrate, as shown in the top row. The far-field spectrum was calculated from the near-field electric field sampled 100 nm above the top surface of the CR. We separately collected the spectra of the full focal plane, the quarter-plane region, and the far field for a periodic nanopillar positioned at the center of the unit cell and at the center of the quarter region, as illustrated in the second and third rows. Owing to the periodic boundary condition, the full focal plane and far-field spectra are identical in both cases, while the quarter-plane spectra exhibit distinct differences, represented by the blue and orange curves. b Light distributions at the focal plane are shown for three different types of pillars operating individually (i–iii), the summed fields from these individual cases (iv), and the actual fields when the pillars operate collectively (v), mapped at wavelengths of 455 nm, 545 nm, and 635 nm. c Transmission spectra of the blue, green, and red channels for the full four-pillar CR and the superposition of the corresponding single-pillar responses. The solid curves show the simulated spectra of the complete CR, whereas the dashed curves represent the linearly superposed spectra from the isolated pillars. d Comparison of wavelength-dependent energy-flow distributions for individual nanopillars and the full CR. Simulated normalized electric-field magnitude $\mathrm{\mid }\mathrm{E}\mathrm{\mid }/\max \mathrm{\mid }\mathrm{E}\mathrm{\mid }$ overlaid with Poynting-vector streamlines in the selected x-z plane for (i) single-pillar configurations and (ii–iv) the full color-router structure at the three target wavelengths: 455 nm, 545 nm, and 635 nm. The schematics above each panel indicate the corresponding pillar configuration and the target detection quadrant at the focal plane. The red/green/blue outlines denote the nanopillar cross-sections intersected by the selected plane, and the white dashed line marks the unit-cell center in the full-router cases.

Although the refractive-index contrast is modest, each pillar still exhibits a distinct wavelength-dependent field confinement and scattering response (Supplementary Note 4). These responses do not act as dominant isolated phase shifters; rather, each pillar contributes only a partial scattered field. The final routing behavior therefore emerges from the combined propagation and interference of all pillar contributions within the same unit cell. In this sense, the CR should be understood as a distributed weak-scattering system, rather than as a simple collection of independently functioning meta-atoms.

To test this interpretation, we compared the focal-plane intensity distributions generated by representative individual pillars, their linear superposition, and the complete four-pillar CR structure, as shown in Fig. 3b(i–v). The superposed field already reproduces the overall routing tendency, suggesting that the individual pillar responses capture an important part of the wavelength-selective redistribution. However, the superposition does not exactly match the actual field of the full CR. The difference indicates that once the pillars are placed within the same unit cell, their scattered fields no longer evolve independently; instead, inter-pillar coupling modifies the local amplitude and phase relations, which in turn reshapes the field evolution during short-distance propagation. Thus, the full response can be understood as an approximate sum of the individual pillar contributions together with a non-negligible correction arising from collective coupling.

To quantitatively evaluate the relationship between the single pillars and the overall CR performance, we calculate the superposition spectrum of the nanopillars:

$${\mathrm{E}}_{\mathrm{full}}\left(\mathrm{x},\mathrm{y;}\mathrm{\lambda }\right)={\sum }_{\mathrm{i}=1}^4{\mathrm{E}}_{\mathrm{i}}\left(\mathrm{x},\mathrm{y;}\mathrm{\lambda }\right)-3{\mathrm{E}}_{\mathrm{bg}}\left(\mathrm{x},\mathrm{y;}\mathrm{\lambda }\right)$$ 2

$${\mathrm{T}}_{\mathrm{superposition}},\mathrm{\lambda }=\frac{{\int }_{{\mathrm{A}}_{\mathrm{\lambda }}}{\left({\mathrm{E}}_{\mathrm{full}}\left(\mathrm{x},\mathrm{y};\mathrm{\lambda }\right)\right)}^2\mathrm{dx}\mathrm{dy}}{{\int }_{{\mathrm{A}}_0}{\mathrm{I}}_{\mathrm{\lambda }}\mathrm{dx}\mathrm{dy}}$$ 3

${\mathrm{E}}_{\mathrm{i}}\left(\mathrm{x},\mathrm{y;}\mathrm{\lambda }\right)$ is the electric field collected at the same focal plane for the single nanopillar, ${\mathrm{E}}_{\mathrm{bg}}\left(\mathrm{x},\mathrm{y;}\mathrm{\lambda }\right)$ is the background field at the detection region when the light is propagated without the CR. The spectra comparison is shown in Fig. 3c. The spectra reconstructed from the superposed single-pillar fields follow the same general trends as those of the full CR, but noticeable deviations remain in both spectral shape and peak value. These discrepancies further confirm that the routing mechanism cannot be reduced to a simple linear addition of isolated responses. Rather, the actual channel performance is governed by collective electromagnetic interactions among the pillars, which determine the balance of constructive and destructive interference in each target region.

A more intuitive picture is provided by the field maps and Poynting-vector streamlines in Fig. 3d. For the complete router, the optical power is progressively reorganized during propagation and redirected toward the designated detection region for each target wavelength. Compared with the isolated-pillar case, the full CR exhibits a more ordered energy-flow pattern and a clearer lateral redistribution of light toward the intended subpixel. This confirms that the routing process originates from cooperative near-field interaction followed by propagation-assisted interference. Even though each LRI nanopillar acts only as a weak scatterer individually, their collective action is sufficient to steer the optical energy over the short propagation distance required by the device.

Therefore, the design principle of the CR is not based on the conventional picture of strong single-element phase control, but on the collective response of multiple weakly scattering low-index nanopillars. PSO is used not as a black-box replacement for physical understanding, but as an efficient tool to identify geometries in which the wavelength-dependent scattering, inter-pillar coupling, and short-range interference cooperate to maximize routing into the target RGB subpixel regions. In this way, all four nanopillars contribute to the redistribution of the three-color channels, allowing the incident light across the full unit cell to be efficiently redirected toward the desired subpixel locations. Although a small amount of residual crosstalk remains, the dominant portion of the optical energy is still routed to the intended regions, consistent with the experimentally observed color reconstruction performance.

Imaging reconstruction

To evaluate the performance of the CR in practical imaging scenarios, we adopted a standardized image reconstruction workflow. A monochrome image was captured through either Bayer CF or our metasurface-based CR to extract RGB channel intensities. Using an RGGB pattern, the signals were demosaiced, white-balanced, and mapped to the CIE color space (see Supplementary Note 12 for details). This work is intended as a proof-of-concept validation of the optical routing mechanism. In a practical implementation, factors such as router-to-sensor spacing, pixel alignment, angular response, and compatibility with the sensor surface stack would need to be carefully optimized, as they can affect routing efficiency and crosstalk. We set up and calibrated the optical path (details in the “Methods” section and Supplementary Note 13) to minimize these influences and avoid mechanically attaching the CR directly onto the imaging sensor; the projected image from a commercial projector (XGIMI Horizon Pro 4 K, 2200 lumens) was first formed at the CR plane using a lens system (details in the “Methods” section: Imaging Characterization). The transmitted light at the focal plane of the CR was then imaged through an objective lens (100× objective lens, 0.9 NA) and recorded by a monochrome camera (BFS-U3-200S6M-C, FLIR). For the CF setup, the CR was removed, and a color CMOS camera (BFS-U3-200S6C-C, FLIR) was used instead.

To demonstrate the color reconstruction capability, five representative color squares-white, green, red, yellow, and blue-were selected. The grayscale images captured prior to reconstruction are shown in Fig. 4a. The redirected light intensities differed across channels, corresponding to varying spectral components that are mapped onto the electrical signals used for color reconstruction (for one unit, there are 22 × 22 pixels used in the monochrome camera). We adopted the sRGB gamut and RGB hex code values to represent the reconstructed color, with detailed values provided in Supplementary Note 11.

Fig. 4. Image reconstruction by integrating a color router in the camera imaging system.

a Single-color square images: white, green, red, yellow, and blue squares imaged through the CR and captured by the monochrome camera. b Reconstructed colors obtained using the color router (CR) and color filters (CF), along with the ground truth (GT), mapped in the CIE coordinate space. c Comparison of pixel brightness under white-light illumination for the CR and CF at varying exposure times. Overexposure occurs when the brightness of the image remains constant with increasing exposure time. d Monochrome camera images of the transmitted focal-plane intensity patterns produced by the CR under progressively reduced white-light illumination, with incident power density decreasing from 312.5 to 12.5 μW/cm². For each illumination level, the upper panel shows the experimentally captured CR period array, and the lower panel shows the corresponding reconstructed color obtained from the different channel intensities (details in Supplementary Note 19). e (i) Ground truth bird image projected through the experimental setup (Supplementary Note 14). Bird images captured using (ii) a color camera and (iii) reconstructed from grayscale images recorded in the CR system. The full reconstructed image was assembled from 17 × 18 blocks, each containing 40 × 40 CR unit cells, raster scanned by shifting the entire image. (iv) A representative monochrome image of a single block and (v) its corresponding reconstructed color image.

We compared the reconstructed colors from the CR-based system with those captured by a color CMOS camera, which uses the same type of sensor as the previously mentioned monochrome type but with an added Bayer filter array in front of the sensor, as shown in Supplementary Note 14. Both sets of reconstructed colors, together with their corresponding ground truth values, were represented in (R, G, B) format and mapped to the CIE 1931 color space for direct visual comparison. Chromatic deviations were quantified by calculating the Euclidean distance between the normalized reconstructed color and the nominal projected color (ground truth) in the CIE space. The ground truth color is directly taken from the projected input. However, slight color changes may occur in the light incident on the CR/CF after transmission through the imaging optics. Since the optical components used in the setup are achromatic, any resulting color mismatch is minor and can be considered negligible.

As shown in Fig. 4b, the CR system exhibits lower chromatic deviations compared to the CF-based system. Specifically, for red, green, and blue, the deviations relative to the ground truth for the CR (CF) were 0.08 (0.20), 0.24 (0.30), and 0.10 (0.15). This CIE analysis indicates that the CR achieves good color fidelity. In typical low-light conditions, color mismatch often arises from crosstalk and reduced light-collection efficiency. Commercial cameras could mitigate these errors by employing longer exposure times or software-based color compensation algorithms. In contrast, our CR system maintains good fidelity without requiring such measures.

To further quantify low-light performance, we disabled the automatic gain function in both systems and manually adjusted the exposure time to capture the white box projection (white color reconstruction). The exposure time was adjusted until both the CF and the CR systems were at the threshold of overexposure, providing a baseline for evaluating the relationship between image brightness and exposure. Overexposure occurs when the brightness of the image remains constant with increasing exposure time. The CR achieved the same brightness at shorter exposure times of 4 ms, while the CF-based system required an exposure time of 10 ms to achieve comparable brightness. As shown in Fig. 4c, as exposure time decreased, image brightness diminished first in the CF system and then in the CR, demonstrating the superior light-collection efficiency of the CR and robustness in low-light environments. Unlike CFs, which lose much of the incident light through filtering, these results show that the CRs were successful in redirecting light instead of absorbing it, thus preserving all spectral components. This effect improves light use, reduces channel overlap, and enhances color fidelity by limiting crosstalk.

Besides, to quantitatively assess the CR response under the same reduced-illumination regime, we measured the monochrome focal-plane intensity patterns at a series of progressively lower incident power densities and reconstructed the corresponding output colors from the recorded channel intensities, as shown in Fig. 4d. In this experiment, the target was a near-achromatic white/gray input, so the evaluation focused on how the reconstructed brightness and chromaticity evolved as the available photon budget decreased. As the illumination decreases, the reconstructed output exhibits a gradual reduction in brightness accompanied by increasing chromaticity drift, which is tolerable in the photon-limited regime.

Building on these advantages, we further validated the CR for full-image reconstruction. Beyond the five representative color squares, the CR system was tested across a broader color spectrum (Supplementary Note 12), including arbitrary colors within the sRGB gamut as well as a more complex subject-a multicolor bird spanning a wide range of hues. Figure 4e shows the reconstructed image, which was assembled from sequential captures of 40 × 40 CR unit-cell blocks and stitched in post-processing (Fig. 4e(iv), (v) show a representative image of one block). For benchmarking, we compared the CR system with a commercial color camera (BFS-U3-200S6C-C, FLIR) using identical pixel sizes and illumination conditions (Fig. 4d(ii)). The CF images appeared darker due to the lower transmission of the unit with filter-based designs, whereas the CR system provided a brighter reconstruction with higher color accuracy. CR image exhibited an average RGB value inaccuracy of ~8% compared to ~20% for CF, and the details are provided in Supplementary Note 16. Even under the same experimental conditions, and despite the wider color range and larger image scale, the CR system captured and reconstructed the bird image with lower error compared to the CF-based system. This result demonstrates the promise of LRI CRs in robust and accurate color separation and color reconstruction.

In summary, these results demonstrate the strong potential of CRs for color imaging, particularly in applications demanding precise color reproduction and reliable performance under low light such as digital photography, industrial inspection, biomedical imaging, and consumer electronics.

Performance analysis

The 2.5D CR implemented in an LRI platform markedly boosts light-collection efficiency in a camera imaging stack. Owing to its gently varying relief and relaxed aspect ratios, the structure is well suited to mass production-for example, via molding and casting into polymer replicas. With improved shape fidelity beyond what we achieved here, we expect further gains in CR performance beyond the ~3× increase in collected photons relative to Bayer filters. Although the proposed LRI CR shows clear advantages in transmission efficiency and low-light operation, several practical limitations remain. The CR maintains stable performance for incidence angles up to 6°. However, for previously reported high-refractive-index periodic CR, performance at relatively large incident angles is also generally limited. This limitation could be mitigated in future LRI CR designs by including multiple incident angles and polarizations in the inverse-design objective, using more complex 2.5D/3D geometries to compensate for angle-induced focal shifts, and co-designing the CR with microlens arrays or pixel-level optical stacks to reduce the effective angular spread. Besides, the current device is also affected by fabrication-related imperfections, such as drift, shrinkage, and field nonuniformity, which can introduce deviations from the intended geometry and reduce performance relative to the ideal design. In addition, because the routing mechanism relies on accurate redistribution of light into designated subpixel regions, crosstalk and SNR must be considered at the system level, as misalignment can degrade both collection efficiency and color fidelity. Residual overlap between routing channels may therefore lead to color reconstruction error, but this can be partially mitigated by mature post-processing methods, such as matrix-based color correction or least-squares compensation, to recover color closer to the ground truth.

To broaden application scenarios and achieve higher performance, we further explored two design variations. First, removing the ITO layer and discretizing the design area into smaller sub-units (unit cells of 40 × 40 pillars, each 50 nm wide with varying heights) yields an overall transmission near 97% while retaining similar per-band peak efficiencies of ~62% (Supplementary Note 3). The key advantage of this variation lies in the increased design freedom afforded by the finer spatial discretization, which enables more effective near-field control and higher-efficiency routing within the low-RI platform. Such flexibility is particularly valuable for engineering advanced color-separation schemes. Second, a low-RI CR with a uniform (single) height achieves competitive performance (Supplementary Note 9) while substantially simplifying tooling and replication for volume manufacturing. Although it sacrifices some light-manipulation ability compared with multilevel designs, the reduced fabrication complexity and tolerance sensitivity make this approach attractive for large-scale industrial deployment, where reproducibility and cost efficiency are critical. These two variations highlight the versatility of the LRI CR platform: one emphasizes design freedom and maximum routing efficiency, while the other prioritizes manufacturability and robustness. Together, they demonstrate how CR architectures can be tailored to balance efficiency, scalability, and cost depending on application requirements.

Discussion

The proposed 2.5D LRI architecture provides a promising route for fabricating high-efficiency CRs for visible-band imaging and sensing, with printed structures that can also serve as reusable molds for scalable, low-cost manufacturing. By broadening the material platform for color routing, this approach takes advantage of the intrinsic benefits of monolithic LRI designs, inherently eliminating interlayer adhesion issues, delamination, and mechanical instability. With appropriate material choices, the concept can be readily extended into the ultraviolet and infrared regimes, broadening its technological reach. In addition, certain LRI glasses exhibit exceptional mechanical durability under extreme environments-including space, deep-sea, high-temperature, and vibration-prone conditions-while their lightweight nature, compatibility, and potential for noise suppression further enhance their utility. By enabling denser photo-pixel integration, the CR can serve as a functional bridge between the optical front end and photodetectors, paving the way for compact, high-resolution, and energy-efficient imaging systems. Looking forward, the direct integration of CRs into imaging sensors represents a natural and important step. Such integration would couple nanophotonic color routing with established semiconductor platforms, potentially improving sensitivity, reducing form factor, and expanding device-level functionality. Beyond color separation, the LRI-based design principle can be generalized across different spectral bands and extended to polarization control or multifunctional photonic operations, thereby opening new opportunities in wavefront shaping, hyperspectral imaging, and advanced optical information processing. We envisage that LRI-based CRs will enable the development of miniaturized, high-efficiency color imaging sensors and unlock new opportunities across digital photography, industrial inspection, biomedical imaging, space exploration, and consumer electronics.

Methods

Finite-difference time domain (FDTD) simulations

Full-wave simulations of the CR were carried out using Lumerical FDTD, while particle swarm optimization was implemented in MATLAB. For the calculation of the near-field transmission objective, periodic boundary conditions were applied in the x and y directions, while perfectly matched layer (PML) boundaries were used along the z-direction. The excitation consisted of x- and y-polarized sources with equal amplitudes to represent unpolarized illumination over the 380–780 nm range.

Fabrication with two-photon polymerization lithography

The fabrication of CR was performed using a two-photon polymerization lithography-based 3D printing technique (Nanoscribe Photonic GT). IP-L was used as the photoresist material for CR fabrication. A 63× objective lens with NA = 1.4 was used in dip-in laser lithography mode. For the fabrication of the CR, the scan mode was still fixed as GalvoScanMode, and GalvoSettlingTime remained at 4 ms. The printing parameters for the CR structures were ScanSpeed was 5000 μm/s, and LaserPower was 55%, corresponding to 27.5 mW. After the exposure process, the sample was developed in propylene glycol monomethyl ether acetate for 5 min, isopropyl alcohol for 5 min, and methoxy nonafluorobutane for 5 min. While immersed in isopropyl alcohol, the sample was UV-cured for 5 min at 70% maximum power (DYMAX MX-150, 405 nm).

Characterization

The CR sample image was taken by a microscope (Nikon Eclipse LV100ND) in transmission mode with a 100× NA 0.9 objective lens. By reducing the aperture of the halogen lamp, the directivity of the broadband white-light source was improved for the imaging test. Clear imaging was obtained by shifting the microscope focal plane from the top surface of the sample to the designed image plane. To evaluate the imaging quality at the wavelengths of 458 nm, 532 nm, and 633 nm, corresponding color filters (FL457.9-10, FL532-10, and FL632.8-10 THORLABS) were placed on top of the halogen lamp, respectively.

Imaging

The projector (XGIMI Horizon Pro 4K 2200 Lumen) was used for the characterization of imaging performance. The optical testing setup is shown in Fig. 4d. The projected image passed through a lens with a focal length of 70 mm, and the light was attenuated using an attenuator. The image was projected onto the CR using a 10× objective lens and then propagated through the device. The transmitted image was magnified by a 100× objective lens, 0.9 NA, and captured by a high-resolution color and monochrome CMOS camera (BFS-U3-200S6C-C and BFS-U3-200S6M-C, FLIR). Both cameras use the same Sony IMX183 1-inch CMOS sensor with a pixel size of 2.40 µm.

Author contributions

C.-F.P., Hao W., Q.R., and J.K.W.Y. conceived the idea. C.-F.P. performed the device design. C.-F.P., Z.L., and X.Z. performed the numerical simulations. C.-F.P., L.C., W.Z., and M.A.R. performed the fabrication. C.-F.P. and T.M. performed the sample characterization. C.-F.P., Hao W., and M.J.H.T. drafted the manuscript. B.W. and Hongtao W. assisted with optical characterization. Hao W., S.Z., M.J.H.T., and Z.L. assisted with schematic drawing. B.W., Hao W, C.-W.Q., and J.K.W.Y. contributed to data analysis and manuscript revision. J.K.W.Y. and C.-W.Q. supervised the project.

Peer review

Peer review information

Nature Communications thanks Hongyoon Kim and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

J.K.W.Y. discloses support for the research of this work from the National Research Foundation (NRF) Singapore, under its Competitive Research Programme award NRF-CRP20-2017-0004, and the National Additive Manufacturing Innovation Cluster (NAMIC), Singapore [H25-MNA0953]. H.W. discloses support for the research of this work from the National Natural Science Foundation of China (NSFC, Grant No. 606HWRC2025117001), the Fundamental Research Funds for the Central Universities (Grant No. 501RCQD2025117002), and the Research Funding of Hangzhou International Innovation Institute, Beihang University (Grant No. 015733207-000001). Q.R. discloses support for the research of this work from the Guangdong Pearl River Talent Program (2023QN10X058) and Shenzhen Fundamental Research Project (GXWD20231130123107001).

Data availability

The data that support the figures and other findings of this study are available from the corresponding authors upon request. Source data are provided with this paper.

Competing interests

The authors declare the following financial competing interests: C.-F.P., Hao W., and J.K.W.Y. are named inventors on a pending patent application related to the color router technology reported in this manuscript. The authors declare no non-financial competing interests. The remaining authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-76072-8.